We used the browser extension tool Image Downloader to retrieve thousands of "space" related images. These were manually curated to remove duplicates, unrelated images, images with non-permissible copyrights, lame ones etc. We normalise the images to be consistently cropped to 512x512 pixels on a black canvas using Image Magick:
j=1;
for i in *;
do val=`printf "%04d" $j`;
echo $i $val $j; convert -background black -gravity center $i -resize 512x512 -extent 512x512 frames_${val}.jpg;
j=$((j+1));
done
Based on their appearance, the images were partitioned approximately into the following categories:
Using these categorised images as input we trained a General Adversarial Network to recoginise the style of images and attempt to produce its own unique spaced-style images. The AI model was run for 2000 epochs, for each epoch the model takes various combinations of the input images and generatively produces its own stylised versions in attempt to pass the image off as an original. At each epoch, up to 128 unique generative assets are produced. We split the individual 512x512 output images up:
convert generated-images-${epoch}.png -crop 514x514 +repage +adjoin tiles_%02d.png
And formulate a network of "similar" looking outputs using findimagedupes:
findimagedupes -t ${chaos} tiles* > tempimg.txt
We then render the final gif of the assets using Image Magick:
convert `cat tempimg.txt` -delay $order -loop 0 final_output.gif;